Summary
The bpftool_maps_access test (converted from shell script to test_progs in commit 2d96bbdfd3b5) fails deterministically on s390x in test_progs_no_alu32 mode and intermittently on aarch64 in test_progs_no_alu32 mode. These failures cause the entire test_progs_no_alu32 CI job to report failure across all independent PRs.
Failure Details
- Test / Component: bpftool_maps_access/unprotected_unpinned (s390x), bpftool_maps_access/nested_maps (aarch64)
- Frequency: Every run on s390x no_alu32; most runs on aarch64 no_alu32
- Failure mode: Wrong result (assertion failure)
- Affected architectures: s390x (unprotected_unpinned), aarch64 (nested_maps)
- CI runs observed:
Root Cause Analysis
s390x: unprotected_unpinned fails at read map btf data
The test calls bpftool btf dump map name not_prot_map and unconditionally asserts success (line 272-273 of bpftool_maps_access.c). On s390x in no_alu32 mode, this command fails with exit code 255.
The bpftool btf dump map command works by:
- Finding the map by name (
map_fd_by_name in tools/bpf/bpftool/common.c:973)
- Getting the map's
btf_id from bpf_map_info (tools/bpf/bpftool/btf.c:948)
- Loading BTF from kernel by that ID (
btf__load_from_kernel_by_id_split at line 1140)
When the map has btf_id == 0 (no BTF associated), step 3 fails. On s390x no_alu32, the kernel apparently does not associate BTF with this map, causing the failure.
The original shell script (test_bpftool_map.sh) was never executed as part of test_progs_no_alu32 — it was a standalone test. The conversion to test_progs (commit 2d96bbdfd3b5) inadvertently exposed it to all test flavors without adding appropriate guards.
Error message: test_basic_access:FAIL:read map btf data unexpected error: 65280 (errno 95)
aarch64: nested_maps fails at create outer map with EEXIST
The test_create_nested_maps subtest calls bpftool map create /sys/fs/bpf/test_bpftool_map/outer_map_tt ... which fails with EEXIST (errno 17), meaning the pin path already exists.
This is a cleanup/isolation issue: if a previous execution was interrupted or if the test_progs retry mechanism re-runs the test, stale pins from the first run may remain.
Error message: test_create_nested_maps:FAIL:create outer map unexpected error: 65280 (errno 17)
Proposed Fix
See 0001-selftests-bpf-Fix-bpftool_maps_access-failures-on-s.patch.
Two changes:
-
BTF guard (s390x fix): Add a map_has_btf() helper that queries the map's kernel info to check if btf_id != 0. Only assert BTF dump success when the map actually has BTF associated.
-
Cleanup before nested map creation (aarch64 fix): Call delete_pinned_map() for both inner and outer maps before creating them, ensuring stale pins from prior interrupted runs don't cause EEXIST.
Impact
Without fix, every s390x test_progs_no_alu32 CI job reports a failure (1 FAILED in Summary), causing misleading CI signal across all PRs. The aarch64 failure similarly causes false positives in the no_alu32 test results.
References
- Commit 2d96bbdfd3b5: "selftests/bpf: convert test_bpftool_map_access.sh into test_progs framework" (Jan 23, 2026)
tools/testing/selftests/bpf/prog_tests/bpftool_maps_access.c:272-273 (BTF dump assertion)tools/bpf/bpftool/btf.c:948,1140 (bpftool BTF dump map implementation)
Summary
The
bpftool_maps_accesstest (converted from shell script to test_progs in commit 2d96bbdfd3b5) fails deterministically on s390x intest_progs_no_alu32mode and intermittently on aarch64 intest_progs_no_alu32mode. These failures cause the entire test_progs_no_alu32 CI job to report failure across all independent PRs.Failure Details
Root Cause Analysis
s390x: unprotected_unpinned fails at
read map btf dataThe test calls
bpftool btf dump map name not_prot_mapand unconditionally asserts success (line 272-273 ofbpftool_maps_access.c). On s390x in no_alu32 mode, this command fails with exit code 255.The bpftool
btf dump mapcommand works by:map_fd_by_nameintools/bpf/bpftool/common.c:973)btf_idfrombpf_map_info(tools/bpf/bpftool/btf.c:948)btf__load_from_kernel_by_id_splitat line 1140)When the map has
btf_id == 0(no BTF associated), step 3 fails. On s390x no_alu32, the kernel apparently does not associate BTF with this map, causing the failure.The original shell script (
test_bpftool_map.sh) was never executed as part oftest_progs_no_alu32— it was a standalone test. The conversion to test_progs (commit 2d96bbdfd3b5) inadvertently exposed it to all test flavors without adding appropriate guards.Error message:
test_basic_access:FAIL:read map btf data unexpected error: 65280 (errno 95)aarch64: nested_maps fails at
create outer mapwith EEXISTThe
test_create_nested_mapssubtest callsbpftool map create /sys/fs/bpf/test_bpftool_map/outer_map_tt ...which fails with EEXIST (errno 17), meaning the pin path already exists.This is a cleanup/isolation issue: if a previous execution was interrupted or if the test_progs retry mechanism re-runs the test, stale pins from the first run may remain.
Error message:
test_create_nested_maps:FAIL:create outer map unexpected error: 65280 (errno 17)Proposed Fix
See
0001-selftests-bpf-Fix-bpftool_maps_access-failures-on-s.patch.Two changes:
BTF guard (s390x fix): Add a
map_has_btf()helper that queries the map's kernel info to check ifbtf_id != 0. Only assert BTF dump success when the map actually has BTF associated.Cleanup before nested map creation (aarch64 fix): Call
delete_pinned_map()for both inner and outer maps before creating them, ensuring stale pins from prior interrupted runs don't cause EEXIST.Impact
Without fix, every s390x
test_progs_no_alu32CI job reports a failure (1 FAILED in Summary), causing misleading CI signal across all PRs. The aarch64 failure similarly causes false positives in the no_alu32 test results.References
tools/testing/selftests/bpf/prog_tests/bpftool_maps_access.c:272-273(BTF dump assertion)tools/bpf/bpftool/btf.c:948,1140(bpftool BTF dump map implementation)